Session: MF-01-01 Application of Fracture Mechanics in Failure Assessment I

Paper Number: 63093

Start Time: Wednesday, July 14, 2021, 09:00 AM

63093 - Machine Learning Modeling of Dynamic Strength of Resistance Spot Welds in High Strength Steels 

Electric resistance spot welding technology has been extensively applied to join sheet metals in the automobile industry to decrease the weight and increase the integrity of motor vehicles. Due to the dynamic nature of motor vehicles, it is necessary to measure and quantify the dynamic strength of spot welds in different metals. This paper will use the traditional regression approach and the state-of-the-art method of marching learning modeling to quantify the dynamic strength of spot welds and then make an evaluation of those two methods through comparisons.

 

This paper collected a large database of experimental results of resistance spot welds with more than 350 tests in three typical steels of high strength steel, dual phase steel and mild steels, where the lap-shear and cross-tension specimens with different thickness over a large spectrum of loading rate, and the strength of spot welds is nominally defined as the peak load achieved during each test that is normalized by the specimen thickness and weld nugget size. The experimental result shows that the strength of spot welds is generally a function of loading rate, specimen thickness, spot weld nugget size, and specimen type. To quantify a general function of the dynamic strength of spot welds, the traditional regression method must make a simplified assumption that the dynamic strength is a function of loading rate only, and then determine a simple analytic function using the curve-fitting approach. However, the machine learning method that adopts a artificial neural network, built-in learning functions and algorithms can determine the dynamic strength as a function of a single variable like loading rate or a function of multiple variables like loading rate, specimen thickness, spot weld nugget size, and so on. To demonstrate the superpower of the machine learning technology, a simple neural network structure with a single input neuron and three hidden neurons as well as a more complex neural network structure with a three-input neurons and a five-hidden neurons are utilized to model the dynamic strength of spot welds. The machine learning results, advantages and disadvantages are finally evaluated in comparison with the regression approach.

Presenting Author: Xian-Kui Zhu Savannah River National Lab

Authors:

Xian-Kui Zhu Savannah River National Lab
Jesse Zhu Cornell University
Wei Zhang The Ohio State University